Inside the Finkel Study Every Matching Algorithm Would Rather You Not Read

In 2012, a team of psychologists led by Northwestern's Eli Finkel published a 60-page review in Psychological Science in the Public Interest that asked a simple question: do the matching algorithms behind online dating actually predict who will be compatible? Their answer, after combing through decades of relationship science, was that matching algorithms have little to no ability to predict real-world romantic outcomes — and the industry has never mounted a convincing rebuttal.

That finding is now over a decade old, but it is more relevant in 2026 than it was the day it published, because the entire swipe-app industry was built on the premise the study dismantles. If you have ever wondered why an app's "99% compatibility" score didn't translate into a good date, this is the paper that explains why — and what would need to be true for a matching system to actually work.

What Finkel et al. Actually Set Out to Measure

The paper, titled "Online Dating: A Critical Analysis From the Perspective of Psychological Science," wasn't a single experiment. It was a systematic review — the authors evaluated the scientific claims dating sites were making about their matching algorithms against the existing body of relationship research, much of it from decades of studies on what actually predicts relationship success.

Their focus was narrow and specific: sites that claimed to use proprietary algorithms — inputting personality traits, values, and preferences — to mathematically predict compatibility between two people who had never met. This is distinct from simple search-and-browse (which just filters by stated preferences) and distinct from communication tools. The question was whether the "matching" step itself was doing anything real.

The Central Finding: Prediction Without Contact Doesn't Work

The core conclusion is blunt: no algorithm can predict compatibility between two strangers with any meaningful accuracy before they've interacted. The traits that are easy to measure and feed into an algorithm — shared interests, personality inventories, stated values — have only a weak relationship to how a relationship actually unfolds once two people are in the same room.

What the research shows matters far more are things that only emerge through interaction: how a couple communicates under stress, how they support each other, whether their attachment styles mesh in practice, and the timing and circumstances surrounding the relationship. None of that is visible from a questionnaire, no matter how long the questionnaire is.

The authors put it in terms that still hold up: matching algorithms can tell you a bit about who someone is, but almost nothing about how a specific pairing will actually feel once it exists.

Why This Isn't the Same as "Matching Doesn't Matter at All"

It's worth being precise here, because the finding gets flattened into "algorithms are useless" more often than it should. The paper doesn't say similarity or shared values are irrelevant — some traits (like a shared desire for children, or similar attachment security) do correlate with better outcomes on average. What it says is that no algorithm can weigh and combine those traits into a reliable individual prediction, because the sheer number of interacting variables in a real relationship overwhelms any model built from profile data alone.

This is a distinction with real consequences. It means the honest use of any pre-interaction signal — human or AI — is as a filter for likely-relevant candidates, not as a guarantee. Anyone claiming their system produces a precise compatibility percentage is making a claim the underlying science doesn't support, regardless of how much data or computing power sits behind it.

What Fourteen Years of Swipe Apps Did With This Finding

The uncomfortable part of this story isn't the 2012 paper — it's what happened after. Swipe-based apps didn't really try to solve the prediction problem Finkel's team identified. They mostly sidestepped it, by replacing "predict compatibility" with "maximize the number of people you see." A deck of thousands of profiles isn't a solution to weak algorithmic prediction; it's an admission that prediction isn't happening, dressed up as abundance.

That shift also happened to be extremely good for engagement metrics. More profiles mean more swipes, more time in-app, and more opportunities to upsell features that promise to cut through the deck you're drowning in. We've written before about why dating apps don't work the way you think they do, and this is the mechanism underneath it: the incentive was never to solve matching, it was to keep the deck moving.

What the Choice-Overload Research Adds to the Picture

Finkel's team also drew on a separate but related body of work: the psychology of choice overload, most famously demonstrated in Sheena Iyengar and Mark Lepper's jam study, where shoppers presented with 24 jam varieties were far less likely to actually buy than those shown just 6. More options didn't produce better decisions — it produced decision paralysis and, often, more regret about whatever was eventually chosen.

Applied to dating, this means the industry's response to "our matching doesn't predict well" — just show people more candidates — actively works against good outcomes twice over. Not only does the algorithm underneath the deck have limited predictive power, but the size of the deck itself degrades the judgment of the person swiping through it. We've gone deeper on this compounding effect in our full breakdown of the math against infinite swiping.

What the Study Does Not Prove

Intellectual honesty requires being clear about the paper's limits, too:

• It doesn't prove that any form of curated introduction is useless — it critiques a specific class of pre-interaction algorithmic prediction, not human matchmaking or hybrid approaches generally.
• It doesn't measure modern large language models, which weren't part of the 2012 landscape. The paper's critique is aimed at fixed-trait questionnaires processed by rigid formulas, not at systems that can hold an open-ended conversation and update their understanding of a person over time.
• It doesn't claim relationship outcomes are unpredictable in principle — only that they can't be predicted well from the kind of static, self-reported data dating sites were collecting.
• It's a review of existing research, not a new controlled experiment against a specific app — its power comes from synthesizing decades of relationship science, not from a single dataset.

None of this rescues the "compatibility percentage" claims that swipe apps still lean on. But it does leave room for an approach that treats matching as a starting point for a real conversation rather than a final verdict — which is a meaningfully different claim than the one the paper knocked down.

Where Agent-Mediated Matching Actually Sits Relative to This

This is where the distinction between a rigid scoring algorithm and a conversational agent matters. Finkel's critique targets systems that reduce a person to a fixed set of checkboxes and run them through a formula once. An AI agent that interviews you, asks follow-up questions, and updates its understanding as you actually go on introductions is doing something structurally different — it's closer to how a skilled human matchmaker builds a mental model of someone over repeated conversations than to a one-time compatibility quiz.

That doesn't mean any AI system automatically escapes the limits the research describes. It means the honest claim to make is narrower and more useful: not "this algorithm knows you'll be compatible," but "this process is built to reduce the number of genuinely mismatched introductions you sit through, and to get better at understanding you as it goes." We go deeper on how that mechanically works in our comparison of what an AI matchmaker actually does differently from a swipe app.

What This Means If You're Deciding How to Spend Your Time

The practical takeaway from over a decade of this research holding up isn't that matching is hopeless. It's that any system — human or algorithmic — that claims to predict compatibility without ever letting the two of you actually talk is overselling itself. The systems worth trusting are the ones that treat their first guess as a hypothesis to test, not a verdict to accept.

A few questions worth asking about any matching approach, given what the research shows:

1. Does it claim a precise compatibility score, or does it explain its reasoning and invite you to weigh in?
2. Does it update based on how your actual introductions go, or is it a one-time questionnaire?
3. Does it reduce the number of people you see, or increase it — and does more volume actually serve you here?
4. Is the incentive behind it aligned with you finding someone, or with you staying engaged?

Services like neverswipe are built around the narrower, research-backed claim: an agent that learns you through real conversation, explains its reasoning, and treats every introduction as new information — rather than a fixed algorithm asked to predict a stranger from a checklist.

Frequently Asked Questions

What is the Finkel study on online dating and matching algorithms?

It's a 2012 review published in Psychological Science in the Public Interest by Eli Finkel and colleagues, evaluating whether dating-site matching algorithms can predict romantic compatibility. The review concluded they have little to no meaningful predictive power over real-world outcomes.

Does this mean matching algorithms are completely useless?

Not entirely. It means static, pre-interaction algorithms can't reliably predict how a specific pairing will feel once two people actually meet. Some broad traits correlate with better outcomes on average, but no formula can combine them into an accurate individual prediction.

Has any newer research overturned the Finkel findings?

No large-scale study has reversed the core conclusion. Subsequent research, including work referenced in Pew Research Center surveys on user satisfaction, continues to be consistent with the idea that algorithmic prediction is limited and that real interaction remains the best test of compatibility.

Why do dating apps still advertise high compatibility percentages?

Compatibility scores are effective marketing even without strong predictive backing. They give users a reason to trust the platform and keep engaging, which serves the app's business model regardless of whether the score reflects genuine predictive accuracy.

How is AI matchmaking different from the algorithms this study critiques?

The study critiques fixed, one-time scoring systems built on static questionnaire data. A conversational AI agent that asks follow-up questions and updates its understanding after each introduction is structurally different — closer to an evolving human matchmaker's judgment than to a one-time compatibility quiz, though it still shouldn't claim certainty it can't back up.